Forensic Analysis of Agentic AI: Lessons from OpenClaw
A new arXiv paper uses the OpenClaw system to establish foundations for investigating agentic AI.
Research Digest
Preprints and papers, rewritten in plain English.
Safety Benchmarks May Measure Policy Labels, Not Detection Ability
A new arXiv paper argues that guardrail classifiers are evaluated against the specific policy they were trained on, not against any universal property of unsafe text.
Shallow Probes Can Guide Personalization Depth in Federated Learning
A new study finds that small client-specific measurements can approximate oracle-level decisions on how deeply to personalize federated models while balancing Macro-F1 against deployment cost.
Error Recognition Boosts Transformer Planning, Study Finds
New research suggests that while planning alone doesn't guarantee error detection, teaching Transformers to recognize errors can improve their planning capacity.
Interpreting the latent worlds of joint-embedding predictive models
A new paper argues that as JEPAs become world models, interpretability must shift from decoding representations to understanding the latent worlds themselves.
Structured RBF Kernels Make Gaussian Processes Practical for Images
A new kernel design for Gaussian processes shares hyperparameters across image channels to handle high-dimensional inputs.
Dynamics-Aligned Representations Tackle Partial Observability in Model-Based RL
A new arXiv preprint proposes aligning learned latent dynamics with the true state to improve model-based reinforcement learning under induced partial observability.
Prompt Injection Detection Needs the Reasoning Window
A new arXiv paper argues that runtime monitors and single-run benchmarks share a flawed assumption about when LLM agents decide to follow prompt injections.
DYNARTmo Adds Acoustics and Aerodynamics to Articulatory Speech Synthesis
An extension to the DYNARTmo articulatory model brings acoustic and aerodynamic simulation to gesture-driven speech synthesis.
PV-Decoding aims to cut inference cost of discrete diffusion language models
A new arXiv paper proposes progressive verification with decode gating to make discrete diffusion language models more efficient.
Diffusion LM sampler timing skews who appears in generated text
New research identifies a second, overlooked decision in diffusion language models—when to unmask a token—as a source of representation bias.
BanglaBBQ: A Bias Benchmark for Bengali Question Answering
A new benchmark adapts the BBQ bias-evaluation framework to the Bangladeshi context, covering both Bangla and English.
Ask, Don't Guess: Targeted Clarification for Streaming ASR Errors
A new method identifies uncertain spans in streaming speech recognition and prompts the speaker to repeat only those parts.
Table Format Changes Which Relations Language Models Preserve
A new diagnostic shows that the way tables are turned into text affects how well instruction-tuned language models can look up and relate values.
Reinforcement Learning Teaches Language Models to Compose Search Primitives
A new arXiv paper proposes using reinforcement learning to let retrieval-augmented language models compose search primitives rather than rely on fixed pipelines.
Multilingual LLMs Encode Language Features Everywhere, Not Just at Edges
New research challenges the claim that language-specific information in multilingual LLMs lives mainly in the first and last layers.
Does LLM unlearning survive team collaboration?
A new arXiv paper asks whether knowledge removal remains effective when large language models work together in shared teams.
Consistency Training Reduces Language Models' Susceptibility to Cues
Consistency training can stop models from changing answers on uninformative cues, but it does not teach cue selectivity.